Hosted text ranking models

Text ranking models, such as Qwen/Qwen3-Reranker-0.6B and Qwen/Qwen3-Reranker-4B, reorder a set of candidate documents by their relevance to a given query. They solve the problem of improving retrieval precision in search, retrieval-augmented generation (RAG), and recommendation systems. Concrete use cases include ranking product listings for an e-commerce search, surfacing the most relevant legal precedents from a document corpus, or prioritizing support tickets by similarity to a customer issue.

In production, these models are typically deployed as a re-ranker stage: after an initial retrieval step (e.g., BM25 or dense embedding search) returns a broad set of candidates, the ranking model scores and reorders those candidates to present the most relevant results. This two-stage pipeline balances recall and precision. The model is called per query, often via an API, and must handle variable throughput with low latency.

Choosing between models involves a trade-off between size, quality, and speed. Smaller models like the 0.6B parameter variant offer lower latency and higher throughput, making them suitable for high-volume, real-time applications where marginal accuracy loss is acceptable. Larger models like the 4B variant provide higher ranking accuracy but incur greater computational cost and response time. For most call volumes, calling a hosted API beats self-hosting by eliminating infrastructure management, scaling overhead, and the need to provision for peak loads, while offering predictable per-request pricing.